DEEP GENERATIVE MODELS: GENERATING SYNTHETIC ELECTROCARDIOGRAM DATA

• Autor(a) principal: Moimenta, Mário Paraíso
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/160549
• topic: Generative Models; Synthetic Data; Electrocardiogram; Generative Adversarial Networks; Variational Autoencoder; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

In the last years, deep generative models have become a popular research topic in artificial intelligence. These models are used in a wide range of applications, including synthetic data generation, missing data imputation, image manipulation, autonomous driving, automatic text translation and speech synthesis. In this thesis we implement three generator models capable of generating synthetic data of electrocardiograms already annotated with the locations of the main waves. The first model is a Wasserstein generative adversarial network with multi-generator. The second model is also a Wasserstein generative adversarial network but with a packed discriminator. The third model is a regular variational autoencoder. In order to prove the quality of the models, we used the synthetic data and the real data to solve a practical problem and compared the performance. The results show that the three models are capable of generating quality electrocardiogram data, which can replace the real data with only a slight loss in data authenticity. The variational autoencoder model produced the highest quality electrocardiograms, but the worse labels. However, the annotations can easily be improved manually. The two generative adversarial network based models produced electrocardiograms with similar quality.